Doktorandské kolokvium KAI - Branislav Zigo (2.6.2025)
v pondelok 2.6.2025 o 13:10 hod. v miestnosti I/9
Prednášajúci: Branislav Zigo
Názov: Visual Perception for Humanoid Robots
Termín: 2.6.2025, 13:10 hod., I/9
Abstrakt:
The ability of humanoid robots to perceive, interpret, and interact with complex dynamic environments is dependent on the development of robust and intelligent visual perception systems. The presentation offers a survey of the visual perception technologies for humanoid robotics.
We begin with classical techniques such as Visual Simultaneous Localization and Mapping (V-SLAM), which provide foundational capabilities for environment mapping and pose estimation. The discussion then progresses through modern advancements including self-supervised 3D representation learning, voxel-based spatial modeling, and vision-based reinforcement learning (VRL), all of which enhance the robot’s understanding of its 3D surroundings and its decision-making autonomy.
A key distinction between visual perception in humanoid robots and conventional computer vision lies in the embodied, real-time, and action-oriented nature. While standard computer vision systems focus primarily on static tasks such as recognition and classification from fixed sensors, humanoid vision is integrated within a closed-loop sensorimotor architecture. It must function in real-time, fuse data from multiple modalities (e.g. stereo/RGB-D, IMUs, LiDAR), and support continuous state estimation, manipulation, locomotion, and social interaction. The perception system is not merely observational but is tightly coupled with motor control and environmental feedback, demanding higher levels of robustness, 3D spatial understanding, and adaptability.
We tackle innovations in deep learning architectures, multi-view sensor fusion, and intrinsically motivated learning strategies. These advancements allow humanoid platforms to perform more reliably in complex tasks such as object manipulation, whole-body control, navigation, and human-robot interaction (HRI). Furthermore, we examine techniques for improving sample efficiency, sim-to-real transfer, and generalization — the key challenges in deploying learned models from simulation to the real world.